{"id":"75bcb1b2-aba1-4b46-978e-abdc3e84d63b","arxiv_id":"2412.11435","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A virtual try-on model replaces explicit garment warping with flow-infused cross-attention guidance in a Stable Diffusion UNet, reporting SOTA scores on VITON-HD and DressCode.","lead":"This paper presents a diffusion model for virtual try-on that injects a predicted garment deformation flow into the model's attention layers rather than warping the garment image upfront. The authors report better texture and alignment than several published methods on VITON-HD and DressCode, though some metric comparisons depend on inconsistent preprocessing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FIDu SOTA claim rests on a non-default cubic resampling variant; under the paper's own bilinear default, Ours (7.914) does not beat GarDiff (7.89), so the central comparison is invalid.","rationale":"I agree with the reader's verdict: the evaluation protocol for the headline FIDu metric is inconsistent. The paper itself flags the sensitivity of FID to downsampling, yet its boldest FIDu result uses a non-default cubic protocol while all baselines use bilinear. The default bilinear result does not beat GarDiff. This directly undermines the abstract's claim of 'significantly outperform' and the SOTA claim in Table 1. The other metrics (SSIM, PSNR, FIDp) do favor Ours, and the DressCode table appears consistent, so the paper is not completely invalid. With a clean, protocol-matched comparison and corrected text, the contribution may still stand. Therefore the reader's CONDITIONAL verdict is appropriate and no change is needed.","tokens_in":11194,"tokens_out":5147,"duration_ms":42339,"concrete_test":"Obtain or reproduce the VITON-HD evaluation pipeline for GarDiff (public code) and run it with bilinear downsampling (the paper's default); then run Ours under the exact same pipeline. If Ours FIDu is not lower than GarDiff's 7.89, the SOTA FIDu claim fails. As a further check, apply cubic downsampling to all baselines; if GarDiff's FIDu also drops below Ours*, the ranking under a shared protocol is unchanged and the bolded number is misleading.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper explicitly acknowledges (Sec. 4.1) that different downsampling methods significantly change FID scores, and states its default is bilinear interpolation following GP-VTON and D4-VTON. Nevertheless, Table 1 bolds an FIDu of 7.869 for 'Ours*', which the caption defines as cubic downsampling, while all baseline numbers come from bilinear. The default 'Ours' FIDu is 7.914, which is slightly worse than GarDiff's 7.89. Thus the headline claim of state-of-the-art FIDu depends on a protocol that the paper itself says is not its default and that no baseline shares. Section 4.2 compounds this by writing 'best score of 7.869' under the FIDp metric, although 7.869 is the FIDu value. Because FIDu is one of the two headline metrics in the abstract and Table 1, the central claim 'significantly outperform' is not supported for FIDu under the stated evaluation protocol. The issue is internal inconsistency, not a matter of external consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"FIA-VTON proposes a diffusion-based virtual try-on method in which a precomputed dense garment-to-model warp flow, local garment features from a garment UNet, and high-level spatial features from FashionCLIP are jointly injected into the cross-attention layers of a Stable Diffusion inpainting UNet. The flow is added to the query and key features rather than being used to explicitly warp the garment, and the authors argue this implicit guidance is less sensitive to flow estimation errors while still preserving detail. The method is evaluated on VITON-HD and DressCode with paired and unpaired FID, SSIM, PSNR, and LPIPS, and the paper claims state-of-the-art results in both datasets.","tokens_in":11441,"tokens_out":5264,"duration_ms":43388,"significance":"The core idea is timely and plausible: conditioning a diffusion try-on model with a dense warp flow through attention, rather than through an explicit warped garment, is a meaningful design choice and the ablation study supports the contribution of both the Flow Guider and the Spatial Guider. If the quantitative claims were robust, the paper would provide a useful alternative to explicit-warping pipelines. The authors also deserve credit for explicitly acknowledging in Section 4.1 that the image downsampling method has a significant impact on FID scores. However, the headline SOTA claim is not currently supported in a load-bearing way: on VITON-HD, the default bilinear model has FIDu 7.914, which is worse than GarDiff's 7.89, and LPIPS 0.047 is worse than GarDiff's 0.036. The bolded FIDu value 7.869 comes from a cubic-downsampling variant, a protocol not shared by any baseline. The DressCode results are more consistently in the paper's favor, but the same protocol concern applies there because the baselines are quoted rather than recomputed. The contribution is worth publishing only after the evaluation protocol is fixed or the claims are recalibrated.","major_comments":[{"comment":"The headline FIDu comparison is not valid under the paper's own protocol. The Table 1 caption states that all numbers use bilinear downsampling except Ours*, which uses cubic downsampling, and Section 4.1 states that different downsampling methods have a significant impact on FID scores. Therefore the bold FIDu value of 7.869 for Ours* cannot be fairly compared with the bilinear baseline FIDu values. Under the default bilinear setting, Ours achieves FIDu 7.914, which is worse than GarDiff's 7.89. The abstract's claim of significantly outperforming state-of-the-art methods is not supported for FIDu on VITON-HD; please recompute all baselines under the same resampling protocol, or remove the SOTA claim for this metric.","section":"§4.1, Table 1"},{"comment":"The text says 'our method achieves outstanding performance in the FID p metric, with a best score of 7.869,' but 7.869 is the FIDu value of Ours*, not the FIDp value; the FIDp value of the default Ours is 4.686. Moreover, on VITON-HD the default model's LPIPS is 0.047, worse than GarDiff's 0.036, and its FIDu is not the best; only FIDp and a marginal SSIM difference (0.913 vs. 0.912) favor the method. The claim of consistent, significant superiority on VITON-HD should be softened or substantiated with a direct comparison using the default configuration.","section":"§4.2, VITON-HD paragraph"},{"comment":"For DressCode, the paper does not state which downsampling protocol was used for the reported FID numbers, nor whether the baseline FID scores were recomputed by the authors or quoted from other papers. Given the documented sensitivity of FID to downsampling in Section 4.1, the DressCode comparisons need the same protocol specification before the DressCode SOTA claim can be accepted. Please state the protocol and, ideally, rerun the baselines or report them with the exact same preprocessing.","section":"Table 2 and §4.1"}],"minor_comments":[{"comment":"There are several typos and wording issues, including 'Wrap Guidance' in the Related Work heading, 'LdDI-VTON' in Section 4.2, 'Spatil Guider' in Section 4.3, 'senarios' in the Section 4.4 title, and 'Table Table 1' in Section 4.2.","section":"Throughout"},{"comment":"The table title says 'Ablation study on each module of FIA-Diff', but the method is called FIA-VTON throughout the rest of the paper; please make the names consistent.","section":"Table 3"},{"comment":"Equation (5) contains a trailing comma in the argument list of epsilon_theta: `∥ϵθ(xt, xg, F, S,) − ϵ∥`; this should be cleaned up.","section":"Eq. (5)"},{"comment":"FashionCLIP is used as the Spatial Guider, but no reference or model version is given; please cite the source and specify the exact feature-map resolution used as the 'highest level feature map without visual projection'.","section":"§3.3"},{"comment":"The ablation text lists replacing FashionCLIP with OpenCLIP as one setting, but the body never discusses the corresponding Table 3 row ('w/ openclip'); either add the discussion or remove the row.","section":"§4.3"},{"comment":"The phrase 'There several possible ways' is grammatically incomplete, and the discussion of the sparsity of warp flow could be more precise about how summation is more stable than multiplication for sparse two-dimensional offsets.","section":"§3.1"}],"recommendation":"major_revision","confidential_remarks":"In my view, the core architectural idea is sound and the DressCode results, if verified under a consistent protocol, are strong enough to support publication. The main obstacle is the VITON-HD FIDu SOTA claim, which currently rests on a non-default cubic-downsampling variant whose bolded value is not comparable to the baselines. The authors should either rerun baselines under the cubic protocol or report the default bilinear model without claiming SOTA on FIDu. I would also ask the editor to ensure that the abstract's 'significantly outperform' wording is revised to match the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick take on arXiv:2412.11435 (FIA-VTON). The core idea is genuinely new: instead of warping the garment explicitly or feeding the warped garment to the diffusion model, they take a dense flow map from D4-VTON's DSDM, project it into the query/key space of cross-attention as additive offsets, and combine that with a FashionCLIP spatial stream. The decoupled cross-attention design is clean, and the ablations in Table 3 show both the flow and spatial modules contribute, with the biggest FIDp drop from removing the flow guider. On DressCode, the method beats GarDiff and D4-VTON consistently across all metrics. That's a solid result.\n\nThe soft spot is the VITON-HD headline. The paper itself states (Sec 4.1) that downsampling method significantly affects FID and that bilinear is the default, following GP-VTON and D4-VTON. Yet Table 1 bolds Ours* FIDu=7.869 from a cubic variant, while the default Ours is 7.914, slightly worse than GarDiff's 7.89. So the abstract's 'significantly outperform' claim doesn't hold for the default protocol. Section 4.2 also says 'best score of 7.869' under the FIDp metric, but 7.869 is the FIDu value. That's a mislabeling, likely a typo, but it's sloppy. The paired FIDp and SSIM/PSNR do favor Ours on VITON-HD, so it's not all bad, but the SOTA comparison needs fixing.\n\nAlso, no code or checkpoints are released, and baseline numbers are quoted from papers rather than rerun. That's not unusual for this area, but it means the central claim isn't independently verifiable right now. The frozen DSDM flow guider is a reasonable choice, and the FashionCLIP spatial stream is a nice complement to the Garment Net.\n\nIf I were refereeing, I'd ask for a corrected table with comparable resampling, a clarification of the metric labeling, and a code release. The architecture is worth engaging with; the DressCode results alone would justify that. It deserves a serious referee, not a desk reject. I'd take it to reading group if you want to see how a plausible conditioning mechanism can be undermined by evaluation protocol details.","headline":"Clever flow-infused attention for try-on, with a solid DressCode showing, but the VITON-HD SOTA claim is undercut by a resampling inconsistency.","tokens_in":12006,"tokens_out":2858,"would_cite":false,"duration_ms":24010,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The central claim is that a dense garment-to-model warp flow, injected through cross-attention as a soft offset to diffusion features, gives virtual try-on both accurate pose fitting and fine texture preservation, beating explicit-warping…","keywords":["virtual try-on","diffusion model","implicit warping","flow-infused attention","cross-attention","garment detail preservation","Stable Diffusion","FashionCLIP"],"falsifier":"Rerun every method on VITON-HD under one common downsampling protocol and recompute FIDp and FIDu; if FIA-VTON's unpaired FID is not below GarDiff's 7.89, the headline state-of-the-art claim rests on the resampling difference the paper itself flags.","tokens_in":10990,"feed_emoji":"👗","tokens_out":6905,"duration_ms":59821,"temperature":0.7,"pith_summary":"Image-based virtual try-on faces a trade-off between fitting a garment to a new pose and preserving its original texture. FIA-VTON's proposal is to stop choosing: it computes a dense warp flow between the garment and the target person and injects that flow into a Stable Diffusion inpainting UNet through a Flow Infused Attention module, so the flow guides deformation softly instead of being applied as a hard warp. The paper reports state-of-the-art scores on VITON-HD and DressCode, improving alignment metrics (SSIM, PSNR, LPIPS) and distribution metrics (FID) over warp-based and learning-based baselines. If the claim holds, virtual try-on quality no longer needs to hinge on the precision of an explicit warping module.","feed_headline":"Warp flow as attention tops virtual try-on benchmarks","feed_subtitle":"Feeding the garment-to-pose flow map into cross-attention preserves texture and fits pose better than warping pixels.","key_machinery":"The central object is the Flow Infused Attention (FIA) module, which replaces every cross-attention layer of the denoising UNet. It takes a dense warp flow $F \\in \\mathbb{R}^{h \\times w \\times 2}$, a local garment feature $G$, the upstream model feature $P$, and a FashionCLIP spatial embedding $S$. The flow is projected by MLPs and added to $P$ and $G$, the summed features go through joint cross-attention with flow-modulated keys and values, and the result is modulated a second time by cross-attention with $S$. This lets the flow act as a learned offset on features rather than a hard warp on pixels, so the diffusion model can correct flow errors while still following the deformation pattern.","core_discovery":"FIA-VTON claims that the dense warp flow from garment to target model, when projected into the feature space of a Stable Diffusion inpainting UNet and injected through cross-attention together with local garment features and high-level FashionCLIP spatial features, guides the model to deform garments to the target pose while preserving fine texture. The paper reports that this implicit flow guidance outperforms explicit-warping pipelines such as VITON-HD, HR-VTON, GP-VTON, DCI-VTON, and D4-VTON, and learning-based diffusion methods such as LaDI-VTON, StableVITON, IDM-VTON, and GarDiff, with the largest gains on paired FID and SSIM. Ablations show that removing either the Flow Guider or the Spatial Guider degrades every metric, and that replacing the flow-attention fusion with concatenation or pixel-wise multiplication of the warped garment harms detail and fit.","pith_inferences":["The paper's headline unpaired FID on VITON-HD (7.869) is computed with cubic downsampling, while the baselines and the default model use bilinear; because the paper itself says downsampling changes FID substantially, a reader should treat the bilinear default (7.914) as the comparable number, and that default is close to GarDiff's 7.89.","Making the flow estimator trainable end-to-end with FIA is a natural next step; since flow quality is load-bearing for detail reconstruction, joint training could produce larger gains than fixing the DSDM flow module.","The flow-as-soft-offset idea should transfer to other image editing tasks with known correspondence fields, such as pose transfer, expression reenactment, or relighting, where explicit warping is also brittle."],"forward_implications":["Virtual try-on quality becomes less dependent on the exact accuracy of the warping module, because flow errors become soft corrections instead of hard pixel distortions.","The warped-garment plus inpainting pipeline can be replaced by a single diffusion UNet that receives flow, local garment features, and spatial features through one attention mechanism.","DressCode results across dresses, lower garments, and upper garments indicate the mechanism transfers across garment categories and handles semi-transparent materials.","Because the Flow Guider is frozen during training, FIA learns to exploit an off-the-shelf dense flow estimator without additional flow supervision.","Flow-guided implicit attention could be adopted by other diffusion-based editing tasks where a correspondence field between two images is known."],"supporting_citations":[{"why":"Supplies the pretrained Stable Diffusion v2.1 inpainting checkpoint that initializes the denoising UNet and Garment Net.","marker":"[1]"},{"why":"Provides the latent diffusion inpainting backbone whose cross-attention layers FIA replaces.","marker":"[25]"},{"why":"Provides the frozen Dynamic Semantics Disentangling Module (DSDM) that estimates the dense warp flow F.","marker":"[29]"},{"why":"Supplies the Garment Net design and the two-UNet detail-preservation idea that FIA adapts for local garment features.","marker":"[31]"},{"why":"Is the main learning-based baseline (GarDiff) that concatenates a warped garment into self-attention, against which FIA compares.","marker":"[26]"},{"why":"Is the warping-free learning-based baseline (StableVITON) whose lack of direct warp guidance FIA addresses.","marker":"[17]"},{"why":"Is an explicit-warping diffusion baseline (DCI-VTON) whose dependence on warp quality FIA claims to avoid.","marker":"[10]"},{"why":"Provides the VITON-HD dataset used for the main benchmark and comparisons.","marker":"[4]"},{"why":"Provides the DressCode multi-category benchmark used to test generalization across upper, lower, and dress garments.","marker":"[22]"},{"why":"Documents how downsampling choices affect FID, the issue the paper itself flags in interpreting its FID results.","marker":"[24]"}],"fun_headline_variants":["Flow attention makes virtual try-on fit better","Implicit warp via attention beats explicit warping","Flow-as-attention: new state of the art for try-on","Garment flow guides diffusion for realistic try-on"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The state-of-the-art claim stands on the assumption that FID scores computed with different downsampling choices are comparable, even though the paper reports that bilinear versus cubic downsampling changes FID scores enough to move which variant appears best.","fun_headline_variants_meta":{"raw":{"variants":["Flow attention makes virtual try-on fit better","Implicit warp via attention beats explicit warping","Flow-as-attention: new state of the art for try-on","Garment flow guides diffusion for realistic try-on"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00078,"raw_usage":{"total_tokens":3431,"prompt_tokens":914,"completion_tokens":2517,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":2455}},"tokens_in":530,"tokens_out":2517,"duration_ms":17169,"temperature":1.0,"reasoning_tokens":2455,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:56:28.779819+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun every method on VITON-HD under one common downsampling protocol and recompute FIDp and FIDu; if FIA-VTON's unpaired FID is not below GarDiff's 7.89, the headline state-of-the-art claim rests on the resampling difference the paper itself flags.","supporting_citations":[{"cited_title":"Stable diffusion 2.1","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained Stable Diffusion v2.1 inpainting checkpoint that initializes the denoising UNet and Garment Net."},{"cited_title":"D$^4$-VTON: Dynamic Semantics Disentangling for Differential Diffusion based Virtual Try-On","cited_arxiv_id":"2407.15111","evidence_quote":"Provides the frozen Dynamic Semantics Disentangling Module (DSDM) that estimates the dense warp flow F."},{"cited_title":"Tryondiffusion: A tale of two unets","cited_arxiv_id":null,"evidence_quote":"Supplies the Garment Net design and the two-UNet detail-preservation idea that FIA adapts for local garment features."},{"cited_title":"Stableviton: Learning semantic corre- spondence with latent diffusion model for virtual try-on","cited_arxiv_id":null,"evidence_quote":"Is the warping-free learning-based baseline (StableVITON) whose lack of direct warp guidance FIA addresses."},{"cited_title":"Taming the power of diffusion models for high-quality virtual try-on with appearance flow","cited_arxiv_id":null,"evidence_quote":"Is an explicit-warping diffusion baseline (DCI-VTON) whose dependence on warp quality FIA claims to avoid."},{"cited_title":"Viton-hd: High-resolution virtual try-on via misalignment-aware normalization","cited_arxiv_id":null,"evidence_quote":"Provides the VITON-HD dataset used for the main benchmark and comparisons."},{"cited_title":"Dress code: High- resolution multi-category virtual try-on","cited_arxiv_id":null,"evidence_quote":"Provides the DressCode multi-category benchmark used to test generalization across upper, lower, and dress garments."},{"cited_title":"On aliased resizing and surprising subtleties in gan evaluation","cited_arxiv_id":null,"evidence_quote":"Documents how downsampling choices affect FID, the issue the paper itself flags in interpreting its FID results."}],"review_version":1}